batch: make /batch a conversational door, not a command wrapper

[Jammy2211]

The /batch skill was a thin wrapper — "run batch plan, return the decision" — which is enough to read a proposal and nothing like enough to run a slot from a chat. This adds the body (skills/batch/batch.md, with SKILL.md pointing at it, matching the repo's convention).

Why

The intended interface is a conversation, not a terminal: paste some tasks from the Mind dashboard, describe the batch you want, say go, and later work through what came back. The CLI exposes one verb; a human says three different things.

The three things a human actually says

"What should I run?" → the BatchDecision as written, with the two parts that must survive into the reply: the review-minute total (the budget is the human's hour, not a task count) and the other lane's ready count when this session cannot plan it. An empty batch at the backpressure cap is reported as a finding, not a deadlock.

"Here are the ones I want." → a paste of dashboard paths, or a description. Resolve it, run the constraints over exactly that set, and report per task what was taken and what was rejected in the planner's own words (autonomy supervised — would park at ship, Status: says the work is already done, epic <slug> phase N is not next, …). Never silently substitute different picks; if everything they named is rejected, say so and offer the planner's proposal as an alternative, not as though it were the ask.

Two rejections may not be overruled: an open Blocked-by:, and Unattended: never. Everything else the human may overrule, provided the reply names which rejection is being overruled and why.

"Go." → vitals, the Heart reason set recorded verbatim, the batch record written before any session starts, then one session per member — never shared, because a shared session serialises them and carries one member's context into the next.

The return path, which was missing entirely

Read the PRs failures-first, then decision-taken, then clean. Offer merge / tweak / reject, and on a tweak, draft the follow-up prompt — the human says one line and never writes a prompt file. Then get review-minutes-actual:, asking for it if it is not volunteered, since it is the only calibration the whole estimate ever gets.

Two guards restated where they will be read

• The independent adversary leg must have run, in a session using a different model, before any PR from a batch is reviewed. If it has not, say so rather than reviewing as though it had.
• delivered: is not "green" — a cloud session's green status means it exited without an infrastructure error, not that the task succeeded.

And the rule that outranks convenience, in both files: the human's go in the conversation is what launches a batch. A schedule may carry the timing; it never carries the authority.

Validation

119 lines against the 200-line mandatory-load budget. 691 tests pass, including the skill-install and local-link guards. (Pre-existing and untouched: skills/prm/ at 251 lines is the one skill over budget.)

---

Generated by Claude Code